fix: train -- #4305, #4351, #4353, #4355

[squid-protocol]

Train of four engine fixes: one golden regeneration, one ledger run, one CI run. Each fix branch is code + regression tests, reproduced on main and audited with the kit. #4306 and #4329 were in earlier builds of this train and have been taken out; see the end.

#4305: inline PERFORM forms draw no callee (fix/4305-inline-perform-callees)

• Broken: calls_out took the word after any PERFORM as a callee:
  • PERFORM TEST AFTER|BEFORE ... → TEST
  • PERFORM FOREVER → FOREVER
  • PERFORM WS-N TIMES / PERFORM 3 TIMES → WS-N / 3
  • EXIT PERFORM [CYCLE] → the next word (END-IF, CYCLE)
• Fix (cobol.py calls_out / _calls_out_ignore; the detector drops an empty callee):
  • TEST and FOREVER are ignored;
  • a name followed by TIMES is the loop count, while out-of-line PERFORM PARA-X 3 TIMES keeps PARA-X;
  • EXIT PERFORM [CYCLE] is consumed by its own capture-less alternative.
• Golden effect: Calls Out To in 3 files: CBSA BANKDATA.cbl, zopeneditor SAM1.cbl, gnucobol CBL_OC_DUMP.cob. In each, the phantom TEST callee is gone.
• Ledger: 4 intra-program-edge false positives fixed, <para> -> PERFORM TEST (CBSA BANKDATA, zopeneditor SAM1 / SAM1LIB / multiroot SAM1).
• Estate: H-0006 FAIL → PASS.

#4351: a short member with stray NULs stays source (fix/4351-stray-nuls)

• Broken: aperture Gate 3.1 blocked more than 1 NUL per 1000 characters as binary. A 168-byte PL/I member with 2 NULs at the end of a comment was dropped (NULREST.pli).
• Fix: at most 4 NULs, and at most 4 other C0 control characters, counts as transfer damage in text, whatever the size. A binary carries control bytes beyond its NULs and is still blocked. The old 16-character "binary" test fixture now carries 5 NULs.
• Golden effect: none; no crucible member has a handful of NULs.
• Estate: H-0037 now has 1 check not passing, down from 5: the member is scanned. The one left is a unit-extent convention question, not this issue. The key's NULREST end_line is 5 (the RETURN), and the engine gives 6 (END NULREST;). For CLMPROC, whose nested procedures follow its last statement, the engine and the key agree. For INIT_CLAIM, the key's own end_line includes its END line. Worth a look at the key.

#4353: national / CJK / full-width names in data_moves (fix/4353-data-moves-national)

• Broken: data_moves' word class was ASCII-only, and a name needed an ASCII letter. MOVE '000001' TO 社員コード gave no row, and MOVE SPACE TO X項目 was recorded into X.
• Fix: the word class takes the record reader's NATIONAL / WIDE_DIGITS / WIDE_HYPHENS characters, and a national letter counts as a letter.
• Golden effect: Data Moves in 3 CICS-Cobol files. Names with _ (in NATIONAL's ID_START, and in the record reader's class) are now whole:
  • WS_NAME was recorded as the two operands WS + NAME;
  • STRING WS-STR_IN1 ... INTO WS-STR_OUT was WS-STR, IN1, ....
• Estate: H-0040, H-0041 and H-0046 FAIL → PASS; H-0042 → PASS together with train A's COBOL engine: a DBCS literal VALUE G'…' is recorded as the value G #4354.

#4355: decimal-comma literal in a MOVE (fix/4355-decimal-comma-move, built on #4353)

• Broken: MOVE 0,5 TO T-WERT (2) under DECIMAL-POINT IS COMMA gave no row, because 0,5 was tokenized as 0 , 5.
• Fix: in such a program a digit-comma-digit run is one numeric literal. A separator comma is followed by a space, and a program without the clause is unchanged.
• Golden effect: none.
• Estate: H-0043 FAIL → PASS.

Taken out: #4306 (fix/4306-idms-phantom-units)

In the first build of this train its step moved 38 golden channels across 515 files. The unit diff shows why:

• Correct removals: the 12 IDMS phantoms (IDMS-CONTROL / PROTOCOL in 6 lsp files), plus USER2 Program1-id and RES FILE-CONTROsL.
• Wrong removals: every real paragraph of NIST NC1134.2.cbl (29 units: OPEN-FILES, MAR-TEST-1..15, PASS, FAIL, ...). That source writes PROCEDURE / DIVISION / . on three lines (116-118), so the branch's "outside the PROCEDURE DIVISION" filter never sees the header.

The branch needs the header matched across line breaks, plus a test with the NC1134 shape, before it goes into a train.

Taken out: #4329 (fix/4329-usage-leak)

The second build of this train went red on CICS Crucible: 5 new hc-abend-link/*/java-ported cells failed mvn compile.

• Engine change behind it: COBOL engine: the last DATA DIVISION entry takes its USAGE from the PROCEDURE DIVISION (DISPLAY verb, 'BINARY' literal) -- wrong byte length #4329 is correct. In HCMAIN, 01 WS-CA. + COPY HCLINKCA. is the last entry before the PROCEDURE DIVISION, and its USAGE leaked in from the procedure text as POINTER. With the fix it is None.
• Forge effect: the forge (gitgalaxy/tools/cobol_to_java, outside this lane) now treats WS-CA as a real record. HCSUB's generated handleLink takes HcsubWsCa, named after the caller's record, instead of HcsubDfhcommarea.
• Why it fails: the crucible's hand-written overlay references HcsubDfhcommarea, so the ported project no longer compiles.

This needs a forge-side decision: name the callee's contract DTO after its own DFHCOMMAREA, or re-key the overlay. Blessing the CICS baseline would only hide the failures. The branch is unchanged and ready to rejoin a train once that is settled.

Golden masters (vs a42fc11fe4a1)

full-precision (tests/golden_master_audit): 5 fixture file(s), 5 channel(s), 6 scanned file(s) moved (cobol 6).

channel	files changed	+added	-removed	leaves	sample
10. Mainframe System Facts / Data Moves	3	0	0	12	cobol/CICS-Cobol/CBL0104v01ProcedureDivision.cbl, cobol/CICS-Cobol/CBL0105v01DeclararElementoGrupo.cbl, cobol/CICS-Cobol/CBL0703v01StatementString.cbl
11. Scan DB Snapshot / Calls Out To	3	0	0	6	cobol/cics-banking-sample-application-cbsa/BANKDATA.cbl, cobol/gnucobol/CBL_OC_DUMP.cob, cobol/zopeneditor-sample/SAM1.cbl
12. Scan Output Snapshots / Graph SQLite	0	0	0	4
7. Mainframe Skeleton Completeness / Channels	0	0	0	4
12. Scan Output Snapshots / LLM Brief	0	0	0	2

Most-moved scanned files (channels): cobol/CICS-Cobol/CBL0703v01StatementString.cbl (1), cobol/CICS-Cobol/CBL0104v01ProcedureDivision.cbl (1), cobol/CICS-Cobol/CBL0105v01DeclararElementoGrupo.cbl (1), cobol/zopeneditor-sample/SAM1.cbl (1), cobol/cics-banking-sample-application-cbsa/BANKDATA.cbl (1)

zero-dependency (tests/golden_master_zero_dep_audit): 5 fixture file(s), 5 channel(s), 6 scanned file(s) moved (cobol 6).

channel	files changed	+added	-removed	leaves	sample
10. Mainframe System Facts / Data Moves	3	0	0	12	cobol/CICS-Cobol/CBL0104v01ProcedureDivision.cbl, cobol/CICS-Cobol/CBL0105v01DeclararElementoGrupo.cbl, cobol/CICS-Cobol/CBL0703v01StatementString.cbl
11. Scan DB Snapshot / Calls Out To	3	0	0	6	cobol/cics-banking-sample-application-cbsa/BANKDATA.cbl, cobol/gnucobol/CBL_OC_DUMP.cob, cobol/zopeneditor-sample/SAM1.cbl
12. Scan Output Snapshots / Graph SQLite	0	0	0	4
7. Mainframe Skeleton Completeness / Channels	0	0	0	4
12. Scan Output Snapshots / LLM Brief	0	0	0	2

Most-moved scanned files (channels): cobol/CICS-Cobol/CBL0703v01StatementString.cbl (1), cobol/CICS-Cobol/CBL0104v01ProcedureDivision.cbl (1), cobol/CICS-Cobol/CBL0105v01DeclararElementoGrupo.cbl (1), cobol/zopeneditor-sample/SAM1.cbl (1), cobol/cics-banking-sample-application-cbsa/BANKDATA.cbl (1)

Ground-truth ledger

Ground-truth ledger: 4 mismatch(es) fixed, 0 introduced.

corpus	fixed	introduced	scoreboard
cics-banking-sample-application-cbsa	1	0	intra-program edges [engine] tp 373/got 374/truth 373 -> tp 373/got 373/truth 373
zopeneditor-sample	3	0	intra-program edges [engine] tp 93/got 96/truth 93 -> tp 93/got 93/truth 93

• fixed: cics-banking-sample-application-cbsa :: engine | intra-program edges | src/base/cobol_src/BANKDATA.cbl | A010 -> PERFORM TEST | fp
• fixed: zopeneditor-sample :: engine | intra-program edges | COBOL/SAM1.cbl | 210-PROCESS-ADD-TRAN -> PERFORM TEST | fp
• fixed: zopeneditor-sample :: engine | intra-program edges | COBOL/SAM1LIB.cbl | 210-PROCESS-ADD-TRAN -> PERFORM TEST | fp
• fixed: zopeneditor-sample :: engine | intra-program edges | multiroot/sam/SAM1.cbl | 210-PROCESS-ADD-TRAN -> PERFORM TEST | fp

Estate-crucible horrors

Estate-crucible (a42fc11fe4a1 -> 27eff08c95d9): horrors passing 37/49 -> 43/49; checks pass 835->855, fail 17->18, missing 14->0, phantom 16->8.

horror	issue	before	after	title
H-0006	#4305	FAIL (7 not passing)	PASS	Inline PERFORM headers and EXIT PERFORM read as PERFORM targets
H-0037	#3534	FAIL (5 not passing)	FAIL (1 not passing)	Stray NUL bytes in a source member (a transfer artifact)
H-0040	#3955	FAIL (4 not passing)	PASS	Japanese user-defined words: PROGRAM-ID, data names (leading and trailing Japanese), sections, paragraphs, _ in names
H-0041	#3991	FAIL (2 not passing)	PASS	Full-width hyphens and digits in names; names that must not collide
H-0042	#3816	FAIL (2 not passing)	PASS	DBCS in EBCDIC: SO/SI-delimited words, mixed and G literals, unbalanced and nested shift codes in comments
H-0043	#3942	FAIL (1 not passing)	PASS	DECIMAL-POINT IS COMMA, with VALUE 1,50 in a copybook
H-0046	#3878	FAIL (1 not passing)	PASS	A member downloaded to a PC in Shift-JIS

channel	before (pass/fail/missing/phantom)	after
data_items	280/6/1/0	281/6/0/0
data_moves	80/0/10/4	91/0/0/3
edges	73/0/0/7	79/0/0/0
entry_points	27/0/1/0	28/0/0/0
programs	27/0/1/0	28/0/0/0
units	91/0/1/3	91/1/0/3

Fixes #4305
Fixes #4351
Fixes #4353
Fixes #4355

🤖 Generated with Claude Code

https://claude.ai/code/session_01FNWHcUSn5p7eMxFTsB5rFo

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab